Researchers uncover how metaphors trick AI into bad code

A team of researchers from CSIRO's Data61 and UNSW Sydney has uncovered a subtle vulnerability in large language models (LLMs) that can degrade code generation performance. Their paper, titled 'Metaphor-Induced Algorithmic Steering: Cross-Domain Procedural Transfer in LLM Code Generation' and published on arXiv (2607.28683), demonstrates how metaphorical language in prompts can implicitly transfer inefficient algorithmic patterns from non-programming contexts into coding tasks.

The researchers developed MASC (Metaphorical Algorithmic Steering for Code Generation), a framework that exposes this phenomenon by iteratively refining metaphorical instructions to elicit suboptimal code implementations. Their experiments show that metaphorical instructions can steer models toward exhaustive search, full scans, or repeated reconstructions—techniques that are efficient in their original domains but counterproductive in programming scenarios. The study found that this effect is detectable through behavioral evaluation and manifests as hidden-state shifts toward lower-efficiency procedural behavior prototypes in the model's internal representations.